Two AI models from Google DeepMind, the tech giant’s research lab, have succeeded in solving math problems at the International Mathematical Olympiad 2024, while AI models have so far failed in logical reasoning.

The AlphaProof and AlphaGeometry 2 models solved four of the six problems set at this year’s international competition for high school students, reaching the level of a silver medalist, a “first” according to Google.

In detail, AlphaProof solved two algebra problems and one arithmetic problem, while AlphaGeometry 2 solved one geometry problem.

The 65th edition of the International Mathematical Olympiad was held in the United Kingdom from 11 to 22 July.

This competition, held since 1959, brings together high school students (and sometimes some exceptional students) selected from around 100 countries.

The first version of AlphaGeometry had already managed to solve 25 Olympiad geometry problems from a set of 30 selected exercises, the scientific journal Nature wrote in January.

“These results open up new perspectives in the field of mathematical reasoning and point to a future where mathematicians and Artificial Intelligence will work together to solve complex problems,” Google said in a press release.

Large language models, AI flagships, have a hard time with logic tests, according to a study published in June in the British Royal Society journal Open Science.

She found that OpenAI’s ChatGPT 3.5 and 4, Google’s Bard, Anthropic’s Claude 2, and three versions of Meta’s Llama responded inconsistently and often relied on illogical reasoning.